Summary:
This diff adds a `vanilla_detectron` workflow so we can invoke the
non-elastic detectron code (essentially whatever is run by `train_net.par`)
via the standard fblearner toolchain. (FAIR folks typically use `fry` for this
purpose...)

I moved most of the code from `train_net.py` to `lib/utils/train.py` so it can
be re-used.

Note that as of now the `vanilla_detectron` workflow does not run testing.
That's coming.

Reviewed By: rbgirshick

Differential Revision: D7325559

fbshipit-source-id: 41359b47132bc79d1fcd5c5992a3589b391d149e